Choose one task and one check
Choose something you already understand: rewrite a note, organize a packing list, or explain a short paragraph. Name one useful result you can check against the original information. “Make a friendly message using only these facts” is easier to judge than “make this better.”
Make a five-minute plan on the worksheet below or a piece of paper. Write the job, the result you will check, a fictional or non-sensitive sample, your time and cost limits, and one reason to stop. Planning and trying the tool are separate steps. An unanswered question is a reason to pause, not to guess.
Check what you would share before sending
Use invented details while you learn. Keep the original notes beside you and leave real documents unchanged. Remove names, contact details, account numbers, passwords, and private work before submission; asking a chatbot to remove them after you upload the original has already shared them.
For the AI service you choose, check what receives or retains your messages, what account connections or tools are enabled, the current request or price limits, and how you can keep the useful output or leave the conversation. Reading, editing, and sending are different permissions. A prompt that says “draft only” does not change an app’s access.
The live lesson helper is archived while its service is offline. You can use the saved first-prompt exercise without sending a message to a model.
Give it a small, fictional job
Fictional practice notes: Plant swap at the community garden. Saturday. Bring spare seedlings. The time has not been decided. Your check: use no more than 70 words, keep every supplied fact, ask about the missing time, and add no food, prizes, fees, or other promises.
Turn these fictional notes into a friendly message of no more than 70 words. Use only these facts. Leave missing details as questions. Notes: Plant swap at the community garden; Saturday; bring spare seedlings; the time has not been decided.
The prompt supplies four useful parts: the task, relevant context, boundaries, and format. Those instructions explain what you want; they do not guarantee that the model will follow them. You do not need a complicated role-play introduction.
Set a small boundary, then inspect the answer
Choose ten minutes or less for the trial and a cost limit, including zero. This exercise needs only plain text: do not connect an inbox, upload private files, or buy another tool to finish it. Stop if the next step needs more access or spending than you planned, pricing is unclear, or the time is up. Make a new plan before going further.
Compare the answer with your original notes. Check that it uses no more than 70 words, and check names, numbers, dates, and promises yourself. Deliberately edited teaching example: “Join us at the community garden on Saturday at 2 p.m. Free snacks provided!” The notes support Saturday and the garden, but this sample drops the plant swap and the instruction to bring spare seedlings. It invents 2 p.m. and free snacks, and states a time instead of asking about the missing time. This is an invented example for checking, not a recorded model failure.
A smooth sentence or a model saying “verified” is not enough. Keep missing details as questions. For a factual claim beyond the supplied notes, consult an appropriate original source; a model answer by itself cannot verify current web information.
Make one follow-up count
Only after the answer passes your fact check should you refine its tone or length. Correct missing or invented details against the original notes first. Then name one change and compare the new answer with the same checks.
Make the message warmer and shorter. Keep every fact from the original notes, add no promises, keep asking about the missing time, and use no more than 70 words.
If the assistant has misunderstood, restate the facts and boundaries. A short model conversation can lose earlier details, so supply the small relevant context again if needed. Stay within the same trial limits rather than chasing endless revisions.
Keep, adapt, or skip
Keep: the result met your check within the limits. Record what worked, and choose one small next use. Your review remains part of the job.
Adapt: there is one specific change worth testing. Name it and set a fresh boundary before another trial. Try a saved prompt on a few different fictional cases before relying on it for repeated work.
Skip: the result or the required access does not fit, or doing the task yourself is easier. Record one sentence about what you learned. A useful outcome can be a decision not to use a tool.
This is a practice framework, not a guarantee of accuracy, safety, or savings. For an urgent or consequential decision, use appropriate current evidence and qualified help rather than treating this exercise as advice.
Keep the plan beside you.
Use the one-page worksheet or make the same notes on paper. Plan one task, choose a sample, write your checks and limits, then record whether to keep, adapt, or skip. Nothing written on paper is submitted to this site.
Download AI worksheet (PDF, 1 page)
3,420 bytes · No signup
Try the saved first-prompt exercise · Read the live-chat privacy note · Check what is safe to paste